1017

MULTIPLE CRITERIA DECISION MAKING IN ACCOUNTING EXPERT SYSTEMS

Daniel E. O'leary School of AccountJng

University of Southern Callomla Los Angeles, Callfomia 90089-1421

USA 213-743-4092

Abstract

The purpose of this paper is to analyze the use of multiple

criterion decision making (MCDM) in accounting expert systems

(AES). Focusing on the MCDM nature of expert system evaluation

functions leads towards a partial taxonomy of AES and a further

understanding of "intelligence" in AES. The focus on the MCDM

nature of the evaluation functions also serves to highlight the

limitations of some of the expert system shells that are

inherently designed for single criteria decision making.

1018

MULTIPLE CRITERIA DECISION MAKING IN

ACCOUNTING EXPERT SYSTEMS

"Any idiot system can maximize a single function."

(Paraphrase of a quote from the Dean of the

Columbia Business School, 1975)

1. Introduction -~"-~"-~---~.-----~--

For the past decade or so, extensive general research has

evolved in the area of knowledge-based expert systems. More

recently, accounting researchers have focused on developing

knowledge-based expert systems for accounting applications.

Virtually all those accounting expert systems have used a

single criteria evaluation functions or have no evaluation

function. Since most decision making involves multiple criteria,

the approach of those systems maybe at a disadvantage when

compared to the approach of human decision makers or the problems

maybe highly limited in scope. There are even some that may

argue that these systems are not "intelligent."

The purpose of this paper is to analyze the use of multiple

criteria decision making (MCDM) evaluation functions in

accounting expert systems (AES) , investigate the evaluation

functions in some AES (and their focus on single criterion

decision making (SCDM» and review the potential MCDM methods

which can be used in AES.

In accomplishing this purpose, this paper elicits a partial

taxonomy for AES and moves towards defin'ing "intelligence" in AES

by focusim'

expert sys1"

function a::

expert sys1i

1. 3 The P:!:

This n

artificial

the modes c::

of the AES ..

accounting

section 5 r::

systems. S3

MCDM and SC:

section 7 dl

section 8 s;

2.

Artifi

science aiml

that take i

than comput

Expert

programs th

domain as w

(R"!:-r and F

1019

n.1f

Irch has

More

)ping

Lons.

~ used a

tion

? criteria,

when

he problems

at may

f multiple

n

ation

rion

methods

a partial

nce" in AES

by focusing on the MCOM nature of the evaluation function in

expert systems. The focus on the MCDM nature of the evaluation

function also serves to highlight the limitations of some of the

expert system shells that are inherently designed for SCDM.

1.3 The Plan Of This,Pa~~

This paper proceeds as follows. Section 2 defines

artificial intelligence and expert systems. Section 3 develops

the modes of decision making that form the basis of the analysis

of the AES. section 4 reviews MCDM, the MCOM nature of

accounting decisions and the implementation of MCOM by humans.

section 5 relates MCDM to decision support systems and expert

systems. Section 6 analyzes some previous AES and discusses the

MCDM and SCOM nature of the AES that have been developed.

section 7 develops some extensions of the use of MCDH in AES.

Section 8 summarizes the paper.

2. Artificial Inte~lli.c]~~~_a_l'l<:i~xpert Sys~lll~

Artificial Intelligence (AI) is that part of computer

science aimed at developing computer programs that perform tasks

that take intelligence and which for the moment humans are better

than computers (Barr and Feigenbaum [1982] and Rich [1983]).

Expert Systems (ES) is a branch of AI. ES's are computer

programs that are designed to perform a task in a specific task

domain as well as a human expert would perform the same task

(B~rr and Feigenbaum [1982J and Hayes-Roth et al. [1983]).

1020

Functionally, ES's may have one or more of the following

characteristics. An ES can perform an intellectually demanding

task rather than a mechanical one. The ES may effectively

interact with the user, e.g., the system may request more

information from the user, if necessary. Or conversely, the user

may request a trace of the reasoning of the ES. Although these

functional characteristics are implemented in many ES's, none of

them is necessary for an ES.

2.1 ES structure

structurally, ES's usually consist of three major

components: database, knowledge base and inference engine. The

database contains the data used by the ES.

The knowledge base contains the knowledge that the ES uses

to process the data. Typically, this is the domain-specific

knowledge that an expert would use to solve the problem.

Generally, this knowledge is symbolic rather than numeric, e.g.,

natural language.

There are a number of different ways of representing

knowledge. Two of the primary methods are rule-based and

frame-based knowledge representation .. The rule-based form of

knowledge representation generally takes the form of "if .•.

(condition) then ... (consequencejgoaljsubgoal)." Often there is

a weight associated with the probability or strength of a rule.

The frame-based form of knowledge representation uses a "frame"

to capture the characteristics associated with a given entity.

The charac:

of interes;

The L

process thl

processes

of alterna!

For exampL

uses eithel

alternatiVi

goal. Whe

determines

also uses

probabilitt

ES's

ES shell.

processin~

of the pr:ii

1984]) an

language c

computer.

states. II

developmerr

knowledge

Two c

and Short]

(1981)).

1021

owing

ly demanding

,t more

'sely, the user

though these

ES's, none of

jor

engine. The

the ES uses

,-specific

blem.

umeric, e.g.,

enting

ed and

ed form of

of "if. ..

ften there is

h of a rule.

as a "frame"

ven entity.

The characteristics define the knowledge about the entity that is

of interest in the application.

The inference engine is the approach used in the program to

process the knowledge base. There are usually at least two

processes represented in the inference engine: choosing the set

of alternatives for evaluation and evaluating the alternatives.

For example, in rule-based systems the inference engine typically

uses either forward or backward chaining to find the feasible

alternatives. Forward chaining is a way of reasoning toward a

goal. Whereas, backward chaining starts with the goal and

determines the approach necessary to accomplish the goal. It

also uses the weights on the rules to evaluate the strength or

probability of the sequence of rules.

ES's usually are developed using either an AI language or an

ES shell. An AI language ~s a computer language that is aimed at

processing symbolic information, such as natural language. Two

of the primary AI languages are Prolog (Clocksin and Mellish

[1984]) and Lisp (Whinston and Horn [1981]). Prolog is the

language chosen by the Japanese for the fifth generation

computer. Lisp has received extensive use in AI in the United

states. An ES shell is software that is designed to aid the

development of ES's. Typically, an ES shell makes the storage of

knowledge easier and prespecifies the inference engine.

Two of the better known ES shells are EMYCIN (e.g., Buchanan

and Shortliffe (1984)) and ALjX (e.g., Duda and Gaschnig

(1981)). Both of these ES shells are rule-based and both use a

1022

weight on the rule to determine the strength or probability of

the 9hosen sequences of rules.

2.2 DecAsion~upport Systems

Decision Support Systems (DSS) are computer-based systems

that are used to (Keen and Scott Morton [1978, p.lJ):

Assist managers in their decision processes in

semistructured tasks.

2. Support, rather than replace, managerial judgement.

3. Improve the effectiveness of decision making rather

than its efficiency.

Some authors imply that ES's are a subset of DSS's (e.g.,

Keen and Scott Morton [1978J), while others argue that they are

not (e'9" Turban and Watkins [1984J). That discussion is

outside the scope of this paper.

2.3 Accounting Expert Systems

AES are expert systems developed to solve accounting-based

problems. There apparently is only one AI-based system that wil

be in commercial use in accounting (willingham and wright

(1985». That proprietary system was developed to examine the

collectability of term and collateral loans. The system was

built using an expert system shell (Insight 2) and has over 1000

rules. The other AES that have been developed are prototypes.

Those prototype systems are the primary focus of this paper.

These systems are discussed in section 6.

.,-' . Modes c

Since E

implies that

processes.

characterizE

judgement, ~

summarized

(Descriptiol

Alternati.

Source: J. D:

The gr:

whereas the

"intelligen:

those two e

Comput:

or symbolic:

computation:

programming;

computation:

is a dictic

backward cr.

1 1023 I

ability of

ed systems

s in

udgement.

ng rather

S's (e.g.,

,at they are

,ion is

Lnting-based

;tem that will

Iright

!xamine the

'stem was

las over 1000

,rototypes.

.s paper.

3. Modes of Decision Making

Since ES perform decision making tasks of experts this

implies that it is critical to understand decision making

processes. In one theoretical approach, Thompson (1964)

characterized four modes of decision making: computation,

judgement, compromise and inspiration. These modes are

summarized in table 1.

Table 1

Thompson's Modes of Decision Making

(Criteria of Choice) Single ~ult~~

(Description of certain-'--­ Computation Compromise

Alternatives) uncertain Judgement Inspiration

Source: J.D. Thompson (1964) as summarized in Zeleny (1982).

The greatest "intelligence" is required in "inspiration,"

whereas there is some question if "computation" requires

"intelligence." "Compromise" and "Judgement" lie somewhere between

those two extremes.

Computation is the typical mode of a highly structured numeric

or symbolic problem. Numeric computation is the best known

computation process. For example, solving the typical linear

programming problem generally requires only computation. Symbolic

computation can arise in many situations. Probably the most basic

is a dictionary table look-up. However, a simple forward or

backward chaining on a set of rules also is computation.

1024

Judgement is the dominant concern of multiattribute utility

theory. Judgement is usually defined by a single dimensional,

clearly stated, but poorly measurable objective (Zeleny [1982]).

Typical objectives include maximizing utility, minimizing employee

dissatisfaction or maximizing the strength of a relationship

between a set of rules.

Compromise is an MCDM process. Compromise involves trading

off on competing objectives, for example a cost:time trade-off.

The first step in compromise is to identify a set of alternatives

and the second step is to reduce this set of alternatives.

Inspiration is the mode of decision making that is commonly

attributed to executives. This approach includes both the

processes of judgement and compromise.

4. Multiple Criteria Decision Making

As noted by Zeleny (1982, p. 1), lIMultiple objectives are all

around us." We may search for the shortest route home, but also

want the least expensive, the most scenic or the fastest route.

Virtually all decision makipg situations include multiple

criteria/objectives. In addition, the objectives may be

conflicting. The Dean of the Columbia Business School has said:

As for conflicting objectives-quality vs. lower cost, better product vs. cheaper raw materials, for example-just about any idiot can maximize a single function. Anybody can increase sales. After all, if nothing else matters, you can decrease the price to zero. In fact, you don't have to stop there. If the! won't take it at zero, you pay them to take it.

Zeleny (

makers may bel

dimension. H[

primarily undl

crisis.

Otherwis.

occurs unless.

"If only one

suffice for mt

there is no "'

considered.

that an "intel

choice criter'

4.1 MCDM and

Before Pl:

question liAr',

There are a n11

One apprct

there are any

There are a nr

(1965) and Lir

Another I

accountants ml

One way of cll

disciplines 0:

1025

lttribute utility

1e dimensional,

(Zeleny [1982]).

Dinimizing employee

relationship

involves trading

:time trade-off.

It of alternatives

:ernatives.

that is commonly

!s both the

objectives are all

:e home, but also

I fastest route.

I multiple

IS may be

School has said:

lower cost,for a single

.fter all, if Ie price to lere. If to take

Zeleny (1982, p. 23) concedes that in some cases decision

makers may be able to express their preferences along a single

dimension. However, Zeleny (1982) suggests that seDM occurs

primarily under extreme conditions of time pressure, emergency or

crisis.

otherwise, Zeleny (1982) contends that no decision making

occurs unless there are at least two decision making criteria.

"If only one criterion exists, mere measurement and search

suffice for making a choice (Zeleny [1982, p.74])." That is,

there is no "intelligence" required unless multiple criteria are

considered. Accordingly, it appears that Zeleny would contend

that an "intelligent" expert system would have at least two

choice criteria.

4.1 MeDM and Accountinq Problems

Before promulgating MeDM for AES, we need to consider the

question "Are accounting problems suitable for MeDM analysis?"

There are a number of ways to answer this question.

One approach is to examine the literature to determine if

there are any papers that relate MeDM and accounting problems.

There are a number of applications summarized in, e.g., Ijiri

(1965) and Lin (1979).

Another approach is to examine the decisions that

accountants make to determine if they have multiple criteria.

One way of classifying these decisions is based on the sub­

disciplines of accounting, e.g., tax and management accounting.

= 1026

Tax accountants face multiple criterion in a variety of heuristics to

environments. Zeleny (1982, p.2) summarizes some of the typical terms of the J

criteria for judging a good tax-shelter: measures may

A second 1. Current deductions from taxable income. 2. Future deductions from taxable income. procedure. Tl 3. Capital gain after selling the investment.

multiple obje,Management accountants provide management with information

[1982]). The: to meet their decision making needs. As noted by Drucker (1974,

solutions to p. 100), management needs to consider the multiple goals of the

find only a "' business:

Much of today's lively discussion of managementby objectives is concerned with the search for "one 5. The RelaJ right objective." This search is not only likely to be as unproductive as the quest for the since ma philosopher's stone: it does harm and misdirects.

To manage a business is to balance a variety of relationshipneeds and goals. And this requires multiple objectives.

5 . 1 MCDM and Accordingly, since management have multiple objectives, so must

Keen ant management accountants.

marriage bet" Auditing also uses multiple objectives. There are a number

intellectualI of criteria in auditing including, e.g., error rate, materiality,

research inta etc.

(1985), but I This discussion suggests that rather than justifying the use

of MCDM in accounting, the problem should be reversed. If SCDM 5 . 2 MCDM am!

is proposed for use in accounting then it should be justified. There hi!!

However, sin« 4.2 MCDM Approaches

also use mul11 There are a number of approaches used by humans to implement

it is anticin MCDM. Probably the most frequent approach is to use a set of

1027

variety of

of the typical

nt.

t:.h information

Drucker (1974,

e goals of the

management 11 for "one 'f likely to

directs. variety of pIe

tives, so must

re are a number

te, materiality,

stifying the use

rsed. If SCDM

be justified.

ans to implement

use a set of

heuristics to aid in the evaluation of a set of alternatives. In

terms of the route home example, the characteristics and their

measures may include distance, time and toll charges.

A second alternative is the use of an MCDM analytic

procedure. These procedures include goal programming, linear

multiple objective programming and compromise programming (Zeleny

[1982). These techniques have the advantage of finding optimal

solutions to the specified problem, whereas, heuristic approaches

find only a "satisfactory" solution.

5. The Relationship of MCDM to DSSand ESls

Since most decision processes are of an MCDM nature, their

relationship to DSS and ES is analyzed.

5.1 MCDM and Decision Support Systems

Keen and Scott Morton~(1978, p.48) indicate that" a

marriage between MCDM and DSS promises to be practically and

intellectually fruitful." There has been general systems

research into this relationship (e.g., Haimes and Chankong

(1985), but little work in accounting MCDM DSS.

5.2 MCDM and ES

There has been little research into the use of MCDM in ES.

However, since many decision processes that require expertise

also use multiple dimensions for the evaluation of alternatives,

it is anticipated that MCDM will have a major impact on ES.

1028

5.2.1 MCDM and ES Shells: SCDM

One of the potential limitations to the use of MCDM in ES is

due to a limitation of some of the ES shells: the shells are

inherently SCDM-oriented. For example, both of the rule-based

systems EMYCIN and AL/X use a single dimension to evaluate the

sequences of rules in the solutions found by the systems. EMYCIN

ranks the alternatives by probability and AL/X ranks the

alternatives by the strength of the relations. For those

situations where a single dimension is appropriate there is no

problem. However, this can lead to making the tool fit the

situation and using an SCDM tool in an MCDM situation.

5.2.2 MCDM and ES: No Evaluation Function

A more extreme situation is the case where there is no

evaluation function. This situation derives from the system

using the symbolic processing nature of the AI language or the ES

shell to "compute" the alternatives and not evaluating the

quality or the feasibilty of the alternatives. This is analogous

to providing the set of constraints to a linear programming

problem and not having an objective(s); that is, this is

computation and not judgement.

5.3 The Use of MCDM in a Taxonomy of AES

Each AES will either have an evaluation function or it will

not have an evaluation function to analyze the quality of the

solutions. If:

process will el

If the sy~

suggests that .

authors sugges~

evaluation fum

evaluation fum

SCDM) serve tOI

systems. This

part of a taxol:

6. Accountin(!

This sect:.

prototype AES III

summarized in 11

(Description 011

Alternatives))

1029

)f MCDM in ES is

shells are

le rule-based

evaluate the

systems. EMYCIN

lks the

)r those

e there is no

01 fit the

tion.

here is no

l the system

lnguage or the ES

lating the

~his is analogous

)rogramming

this is

:tion or it will

lality of the

solutions. If it has an evaluation function process then that

process will either be an MCDM or an SCDM process.

If the system does not have an evaluation function then that

suggests that the system is only a computational system. Some

authors suggest that virtually all decision problems have an MCDM

evaluation function. Accordingly, the existence of the

evaluation function and the type of evaluation functions (MCDM or

SCDM) serve to distinguish between characteristics of alternative

systems. This indicates that these two characteristics can be

part of a taxonomy of AES.

6. Accounting Expert Systems

This section summarizes and categorizes the existing

prototype AES's into the categories of table 1. The results are

summarized in table 2.

Table 2

AES by Mode of Decision Making

(criteria of Choice) Single Multiple

(Description of certain TAXADVISOR FINSTA

Alternatives) uncertain AUDITOR EDP AUDITOR

N/A

1030

6.1 TAXADVISOR

One of the best known AESls is TAXADVISOR. TAXADVISOR is a

rule-based system that was built using EMYCIN. However, the

system does not use the certainty factor capabilities inherent in

EMYCIN.

TAXADVISOR (Michaelsen [1982]), is an AES designed to make

recommendations concerning estate planning. The system

recommends a set of actions based on a set of conditions.

The system functions as an reference source. The system

provides the user with a "dump" of its knowledge base given a set

of conditions. The resulting set of actions may be economically

feasible or the set of actions may be economically infeasible or

the set of actions may contain contradictory recommendations

(Michaelsen [1982, pp. 166-170]).

Accordingly, one of the major limitations of TAXADVISOR is

that it does not know if the actions it recommends are

economically feasible or if the actions are contradictory. The

system has no evaluation function to analyze the recommendations.

Accordingly, this system is categorized in the "computation"

decision mode of table 1.

The system could be extended to include the use of the

certainty factor. However, as noted above, a ranking of actions

based on the certainty factor is an SCDM process.

6.2 AUDI'lI

AUDI'lI

built usirr

judgements

debts.

EDP 1+\

system buji

the audit

Since

built usirn

function.

derives a

solution.

devise. AI

II judgement

6.3 FINST

FINST

the desigru

aggregatedl

to improvel

model the

The .i.

needs of t:

set of alt:

second pha

1031

TAXADVISOR is a

However, the

lities inherent in

designed to make

e system

onditions.

e. The system

e base given a set

y be economically

lly infeasible or

commendations

of TAXADVISOR is

:nds are

Itradictory. The

,e recommendations.

"computation"

,e use of the

'anking of actions

,so

6.2 AUDITOR and EDP AUDITOR

AUDITOR (Dungan [1983]) is a rule-based system that was

built using AL/X. AUDITOR was developed to make diagnostic

judgements concerning the adequacy of a firm's allowance for bad

debts.

EDP AUDITOR (Hansen and Messier [1985]) also is a rule-based

system built using AL/X. EDP AUDITOR was developed to assist in

the audit of computerized accounting systems.

since both AUDITOR and EDP AUDITOR are rule-based systems

built using AL/X, both of the systems use the AL/X evaluation

function. This function measures the strength of a rule and

derives a single "strength" measure associated with each

solution. These SCDM systems each make use of this probabilistic

devise. Accordingly, these systems are categorized in the

"judgement" section.

6.3 FINSTA

FINSTA (Munakata and O'Leary [1985]) was developed to aid in

the design of accounting information systems by developing

aggregated financial statements from a set of accounts in order

to improve management decision making. The system uses Prolog to

model the judgements of a management consultant.

The inference engine was developed to meet the specific

needs of the problem. The first phase of the program developed a

set of alternative aggregations of the financial statements. The

second phase chooses from among those alternatives based on two

1032

different criteria. Since this system uses multiple criteria but

does not use probability its mode of decision making is

categorized as "compromise."

7. Extensions of MCDM in AES

If an ES shell is used then the inference engine is

prespecified. Accordingly, in those cases the evaluation process

is a function of the inference engine. As a result, whether an

SCDM or an MCDM process is used is dependent on the ES shell's

inference engine. If the evaluation process is an MCDM process,

then the user must ensure the ES shell has an MCDM evaluation

process.

However, if an AI language is used to develop the system

then the knowledge base can be interfaced with an arbitrary MCDM

process. In particular, the inference engine can use the

knowledge base to develop the feasible set of alternatives or the

feasible region for MCDM evaluation. For example, in FINSTA the

inference engine used the knowledge base to develop the

alternatives that were later evaluated using a heurIstic approach

that was based on two different criteria. Another alternative is

to use an MCDM optimization method (e.g., linear multiobjective

programming or goal programming) to generate an optimal solution

to the established problem.

If a technique such as goal programming is used then there

is another step that can be taken: An expert system can be

developed to analyze the dual variables, etc. Understanding the

output is an

problem, SiD

of mathemat:

of developil

mathematica:

(1985).

8. Summar

This p)

functions i.

importance

some protot:

MCDM evalual

relating tc:

This fI

feasible (a:

MCDM AES's ..

valuable cH

in the chall

" intelligell

judgement,

do not app"

many of th(1

formulatio):

1033

Ltiple criteria but

aaking is

engine is

evaluation process

asult, whether an

~ the ES shell's

I an MCDM process,

~CDM evaluation

alop the system

an arbitrary MCDM

::an use the

!lternatives or the

~le, in FINSTA the

{elop the

heuristic approach

cher alternative is

~r multiobjective

1 optimal solution

I used then there

rstem can be

Understanding the

output is an important part of any mathematical programming

problem, since, as noted by Geoffrion (1976,p.444), "the purpose

of mathematical programming is insight not numbers." The process

of developing an expert system for use in analyzing a

mathematical program is discussed in more detail in O'Leary

(1985).

8. Summary

This paper has focused on the nature of the evaluation

functions in AES. This paper is the first to suggest the

importance of MCDM in AES. This is substantiated by a review of

some prototype AES's that found only one system that employs an

MCDM evaluation function. There was no discussion in that paper

relating to the MCDM nature of the system.

This focus has lead to four primary findings. First, it is

feasible (as substantiated by FINSTA) and desirable to develop

MCDM AES's. Second, the evaluation function appears to be a

valuable characteristic in a taxonomy of AES. This is summarized

in the characterization of the decision mode. Third, an

"intelligent" AES must be in the decision mode of either

jUdgement, compromise or inspiration. Purely computational AES

do not appear to meet the requirements of intelligence. Fourth,

many of the ES shells are of an SCDM nature. This can lead to

formulation of MCDM problems as SCDM problems.

1034

REFERENCES

Barr, A. and Feigenbaum, E.A., The Handbook of Artificial Intelligence, Heuristech Press, Stanford, Ca. and William Kaufman, Los Angeles, Ca., 1981.

Buchanan, B.G. and Shortliffe, E.H., Rule-Based Expert Systems,Addison-Wesley, Reading, Massachusetts, 1984.

Clocksin, W.F. and Mellish, C.S., Programming in Prolog, Springer-Verlagg, New York, 1984.

Drucker, P., Management Tasks, Responsibilities and Practices, Harper and ROw, 1974.

Duda, R.O. and Gaschnig, J.G., "Knowledge-based Systems come of Age," Byte, September, 1981, pp. 238-28l.

Dungan, C., A Model of an Audit Judgement in the Form of an Expert System, Unpublished Ph. D. Dissertation, University of Illinois, 1983.

Geoffrion, A.M., "The Purpose of Mathematical Programming is Insight, Not Numbers," Interfaces, 1976, Vol. 7, p. 81.

Haimes, Y.Y. and Chankong, V., Decision Making with MultipleObjectives, Springer-Verlag, Berlin, 1985.

Hayes-Roth, F., Waterman, D. and Le~at, D., Building ExpertSystems, Addison-Wesley, Reading, Massachusetts, 1983.

Hansen, J.V. and Messier, W.F., "A Knowledge-Based Expert system for Auditing Advanced Computer Systems," Unpublished WorkingPaper, school of Business, University of Florida, January, 1985.

Ijiri, Y., Management Goals and Accounting for Control, North-Holland Publishing Co., Amsterdam, 1965.

Keen, P. and Scott Morton, M.S., Decision Support Systems, Addison-Wesley, Reading, Massachusetts, 1978.

Lin, T., "Application of Goal Programming in Accounting," Journal of Business Finance and Accounting, vol. 6. no. 4, 1979, pp.559-577.

Michaelsen, R., A Knowledge-Based System for Individual Income and Transfer Tax Planning, Unpublished Ph.D. Dissertation, University of Illinois, 1982.

Munakata, T. and O'Leary, D., "A Prototype Business Expert System: Aggregation in Internal Accounting Systems," UnpublishedWorking Paper, School of Accounting, University of Southern California, December, 1985.

O'Leary, D., presented at Engineering,

Rich, E., Art

Thompson, J.n W.W. cooper,in Organizatil

Turban, E. am Decision supp:: Business, Unil

Willingham, JJ System for AUl paper present:

Winston, P. al

Massachusetts

Zeleny, M., ~ York, 1982.

Hermes,

tificial William

E:xpert Systems,

:ld Practices,

(stems come of

1'orm of an 1iversity of

Jramming is ). 81.

:h Multiple

ling Expert L983.

I Expert Systemled Working January, 1985.

Systems,

mting," Journal 1979,

ridual Income Irtation,

IS ExpertI," Unpublished , Southern

1035

.,I

O'Leary, D., "Expert Systems in Mathematical Programming," paper presented at the Symposium for Artificial Intelligence in Engineering, washington, D.C., October, 1985.

Rich, E., Artificial Intelligence, McGraw-Hill, New York, 1983.

Thompson, J.D., "Decision Making, the Firm and the Market," in W.W. Cooper, H.J. Leavitt and M.W. Shelly (eds), New Perspectives in Organization Research, Wiley, New York, 1964.

Turban, E. and Watkins, P., "Integrating Expert Systems and Decision Support Systems," unpublished Working Paper, School of Business, University of Southern California, 1984.

Willingham, J. and Wright, W., "Development of a Knowledge-based System for Auditing the Collectability of a Commercial Loan," paper presented at the ORSA/TIMS meeting, Boston, 1985.

Winston, P. and Horn, B., LISP, Addison-Wesley, Reading,Massachusetts, 1981. ----

Zeleny, M., Multiple Criteria Decision Making, McGraw-Hill, New York, 1982.

FOOTNOTE

1. Hermes, vol. 3, no. 2, spring-summer 1975.

Sixiemes Journees Internationales LES SYSTEMES EXPERTS &LEURS APPLICATIONS

The 6th International Workshop on EXPERT SYSTEMS &THEIR APPLICATIONS

28-30 avril 1986 - April 28-30 1986 I I I

I

t= I

~l-I-­VOLUME 2